itzune/pocket-tts-basque-onnx

Pocket TTS Basque — ONNX (euskarazko ahots-sintesia, PyTorch gabe)

ONNX export of itzune/pocket-tts-basque, the 6-layer Basque (euskara) Pocket TTS model, for streaming text-to-speech on a CPU with onnxruntime only — no PyTorch. Same weights, same five bundled voices, same zero-shot voice cloning; fp32 and int8 graphs.

The bundle follows the format of KevinAHM/pocket-tts-onnx (the community ONNX export of Kyutai's Pocket TTS), whose reference runtime is included here as pocket_tts_onnx.py.

Files

file	what it is
flow_lm_main[_int8].onnx	transformer backbone (text + voice conditioning, one latent frame per step)
flow_lm_flow[_int8].onnx	flow-matching head (latent sampling)
mimi_encoder[_int8].onnx	Mimi encoder, for cloning a voice from a clip
mimi_decoder[_int8].onnx	Mimi decoder, streaming (latent frames → 24 kHz audio)
text_conditioner[_int8].onnx	token embeddings
bundle.json, bos_before_voice.npy	state layout and metadata for the runtime; learned BOS vector placed before the voice
tokenizer.model, vocab.json, token_scores.json	Basque SentencePiece tokenizer (4000 tokens), also as JSON
voices/*.safetensors	the five bundled voices (Maider, Antton, Nahia, Ane, Aitor; see voices/voices.json)
pocket_tts_onnx.py	reference streaming runtime by KevinAHM (Apache-2.0)

The int8 graphs use dynamic MatMul quantization. The reference runtime always runs the Mimi encoder and the text conditioner in fp32 and uses int8 for the other three, so precision="int8" needs about 160 MB of files and precision="fp32" about 440 MB; the mimi_encoder_int8 and text_conditioner_int8 graphs are there for other runtimes.

Usage

pip install onnxruntime numpy sentencepiece safetensors huggingface_hub soundfile

import sys
import numpy as np
import soundfile as sf
from huggingface_hub import snapshot_download

bundle = snapshot_download("itzune/pocket-tts-basque-onnx", allow_patterns=[
    "*int8.onnx", "mimi_encoder.onnx", "text_conditioner.onnx",  # the runtime runs these two in fp32
    "*.json", "*.npy", "*.model", "*.py", "voices/*"])
sys.path.insert(0, bundle)
from pocket_tts_onnx import PocketTTSOnnx

# temperature 0.3 is this model's default (the runtime's own default, 0.7, is for Kyutai's models)
tts = PocketTTSOnnx(models_dir=bundle, precision="int8", temperature=0.3)

voice = f"{bundle}/voices/Maider.safetensors"   # or a 5-10 s .wav of clean Basque speech to clone
audio = np.concatenate(list(tts.stream("Kaixo! Hau euskarazko ahots sintetiko bat da.", voice)))
sf.write("kaixo.wav", audio, tts.sample_rate)  # 24 kHz

tts.stream() yields audio as it is generated (the first chunk after two latent frames), so playback can start long before the sentence is finished; precision="fp32" uses the full-precision graphs.

Notes:

• Write the text as it should be read: numbers, dates, abbreviations and symbols must be spelled out in words (e.g. hogeita lau rather than 24). The dataset's normaliser (bssd.text_norm) does this for standard Basque.
• Not compatible with sherpa-onnx as is: this model inserts a learned BOS vector before the voice (bos_before_voice.npy), which sherpa-onnx's Pocket TTS implementation does not apply.
• Only clone voices you have the right to use.

Speed

scripts/bench_onnx.py (three sentences of 1-7 s, Maider voice, streaming, generation only, after a warm-up) on an 8-core x86 laptop CPU, with the PyTorch model on the same machine for comparison:

runtime	threads	first audio	speed (× real time)
ONNX int8	2	0.07 s	4.4×
ONNX fp32	2	0.09 s	3.5×
PyTorch fp32	2	0.07 s	6-7×
PyTorch int8 (--quantize)	4	0.03-0.05 s	8×

The ONNX runtime is slower than PyTorch but still several times faster than real time, with first audio in about 0.1 s, at a fraction of the install size (onnxruntime and friends: ~150 MB, against several GB for PyTorch).

Export

Exported with the community exporter lomotron/pocket-tts-onnx-export (pinned commit, with a small fix for exporting from a config file) from the released basque.yaml; every graph was verified against the PyTorch model within tolerance before quantization. Reproduce with scripts/export_onnx.sh in https://github.com/itzune/pocket-tts-basque.

Limitations

As for itzune/pocket-tts-basque: Basque only (code-switched words are read with Basque phonetics); a small share of sentences (about 1-2 %) comes out as a continuation of the reference clip instead of the input text; trained on synthetic speech. int8 quantization was checked by ear and with ASR on the benchmark sentences, not with the full evaluation of the PyTorch model.

License and attribution

Model weights: CC-BY-4.0 (as itzune/pocket-tts-basque and the upstream Pocket TTS weights). pocket_tts_onnx.py: Apache-2.0, from KevinAHM/pocket-tts-onnx. Please cite Kyutai's Pocket TTS, the HiTZ-Aholab studio corpus for the Maider and Antton voices, and the BSSD dataset (itzune/bssd, CC-BY-SA-4.0).